Gemini 3.5 Flash Lite (Vertex AI, EU)

Positioned as a high-efficiency tier within the Gemini line, this Flash Lite variant is built with upgraded agentic capabilities and aimed squarely at sub-agents that handle focused assignments inside broader multi-agent workflows. Its multimodal ingestion spans text, images, video, audio, and PDF, while generation stays text-only, making it suitable as a utility worker in pipelines where a larger model delegates narrow jobs. The EU endpoint designation is consistent with provider-region routing for European deployments and aligns with the platform’s data-governance posture of not retaining inputs or using them for training.

Reported benchmark results for the release highlight balanced reasoning and coding performance for its size class, with a GPQA Diamond score of 83.8% on graduate-level science questions, an Artificial Analysis Intelligence Index of 37.4%, and a Coding Index of 49.3% that combines evaluations like LiveCodeBench, SciCode, and Terminal-Bench. In the Agent Death Trap survival benchmark, this Gemini variant finished a 14-room corridor run on 70 HP, described as ahead of every other Gemini on the board at that update. With support for tool calling, structured output, caching, and web search, alongside a one-million-token context window and 66K-token output ceiling, the model fits cost-sensitive agent loops where repeatable task execution matters more than top-tier reasoning scores.